DeepMind is speeding toward artificial general intelligence (AGI)…
by Jeff Brown, editor, The Bleeding Edge
Google’s artificial intelligence (AI) division DeepMind just came out with another incredible development. It seems like these breakthroughs are coming nearly every month now.
This month DeepMind released a “generalist agent” for the very first time. It’s called Gato. And Gato is one of the world’s first multi-modal, multi-tasking AIs.
Here’s what that means…
We have talked about all kinds of different AIs in The Bleeding Edge. The one thing they all had in common is that they are each optimized for a single task.
For example, DeepMind’s AlphaFold predicts the folding of proteins. Meanwhile, Miso Robotics’ Flippy turns burgers on the grill. And, of course, there are many other AIs trained to master a specific task.
What sets Gato apart is that it mastered many different tasks. In fact, DeepMind trained the AI to do 600 different things.
Here’s a look at just a few of them:
Gato in Action
Source: DeepMind
Here we can see a sample of what Gato can do. It can operate a robotic arm designed to pick, sort, and categorize objects. It can manipulate and organize building blocks. It can notate the objects it “sees” in an image or video. It can play different video games.
Gato is also skilled in natural language processing. It can understand human language and even engage in conversations.
So Gato has learned to do many, if not all, of the tasks that the other AIs can do. That’s remarkable. If we think about it – even humans might struggle to be good at 600 separate things.
And get this… On a relatively small training set, Gato was able to achieve an expert level score on 450 out of the 600 tasks that it learned.
And what’s so exciting is that the number of tasks is arbitrary. If Gato can learn 600 tasks, it’s just a matter of scaling the training for it to master 60,000 tasks.
What we are looking at is the natural evolution of AI. We are moving from single-task AIs to more general, multi-tasking AIs that can do all kinds of things. This is the next logical step toward artificial general intelligence (AGI).
AGI is when an AI is indistinguishable from a human and is empowered with “superhuman” capabilities. It can think, understand, learn, and apply its knowledge to solve any problem – just as we humans do – but in a fraction of the time and on any subject matter at all.
This latest development from DeepMind signals that we are rapidly moving towards that eventuality. It’s a little bit frightening to think about how fast AI is advancing.
The implications are profound. Everything changes when we have AGI. We can’t even predict what the world looks like with AIs that can think and act independently.
How Gato develops over the next 12 months will be very telling.
|
